AI referral traffic is often discussed as a future channel. In one small SaaS property, it is already material.

During the latest measured window, 22.3% of sessions came from AI assistants. Almost all of those visits came from ChatGPT.

This is a first-party GA4 analysis of a single site, not a market benchmark. The absolute base is small. The useful story is the share, its direction, and the behavior of the people who arrived.

The four findings

1. AI assistants referred 22.3% of all sessions

More than one in five sessions in the measured window carried an AI-assistant referral source. This is AI-referred human traffic recorded by GA4.

It is not crawler traffic. Crawlers generally do not execute the analytics JavaScript used to record these sessions, and GA4 cannot be used to measure verified crawler activity.

2. AI traffic was 64% the size of Google organic

The comparison is not “AI versus all search.” It is narrower and more useful: AI referral sessions were 64% of Google organic sessions in the same window.

Google organic remained the larger channel. AI referrals were already large enough that omitting them would leave a substantial hole in an acquisition report.

3. AI visitors engaged above the site average

The AI-referred visitors did not behave like empty clicks. Their engagement rate was above the site average.

That framing matters. Referral sessions show that people arrived; engagement provides evidence that at least some of those visits were relevant. Neither proves that every visit converted, and neither reveals what any crawler did.

4. Weekly share rose from 4.5% to 26.3% in eight weeks

The weekly series moved from 4.5% of sessions to 26.3% over eight weeks. That is the strongest reason to preserve a baseline now: a monthly snapshot alone can hide how quickly the channel is changing.

The correct conclusion is not that every site should expect the same curve. It is that AI referral traffic can become reportable faster than an annual planning cycle.

What the assistants sent people to

The leading landing pages were useful, task-oriented pages rather than a generic sales page. That pattern suggests a practical content lesson: pages that solve a specific problem can become both search assets and AI-recommendation destinations.

This study does not identify why an assistant selected a particular page. GA4 records the referred visit and landing page; it does not expose the answer that generated the click. Connecting those layers requires separate, disclosed prompt sampling.

How the analysis stays honest

The classification begins with a GA4 Traffic acquisition export at the session source / medium level. Known AI-assistant referrers are grouped, then compared with all sessions and Google organic sessions from the same export.

The boundaries are explicit:

  • the data describes human sessions recorded by GA4;
  • it does not measure crawlers;
  • referral data does not capture an AI mention that produced no click;
  • one site is not a universal benchmark; and
  • the absolute base is small, so the share and trend deserve more attention than the volume.

That is enough to answer a valuable question: Are AI assistants already sending people to this site, and is the channel becoming material?

Build the baseline before telling the story

For consultants and small agencies, the immediate workflow is straightforward:

  1. export GA4 Traffic acquisition data using session source / medium;
  2. calculate the share of AI-referred human visits;
  3. compare AI referrals with Google organic in the same window;
  4. compare engagement without calling it conversion; and
  5. save the report so the next window has a trustworthy baseline.

Brandvane’s free AI Traffic Report performs that calculation locally in the browser. The analytics file is never uploaded, transmitted, or stored.

The output should remain modest about what it knows. A referral is evidence that a person arrived. A verified server-log event is evidence that infrastructure fetched a page. An appearance in a sampled answer is evidence about that sample. The value comes from connecting those signals without collapsing them into one inflated claim.